FPGA based Deep Learning Models

Real-time object detection and recognition systems have been of more importance and find heavy applications in various fields. There are different works related to object detection and recognition models. Recent object detection and recognition models are based on deep learning techniques.

Convolutional Neural Networks find good results in terms of dependability on detection in real-time.

The initial model for detection and recognition is done on the YOLO model. There is a high need for faster and reliable object detection and recognition models and systems that can be implemented in real-time. The object detection and recognition models that are previously implemented are without many FPGA-based experiments.

SSD and FRCNN

Single Shot Detector SSD uses CNN’s pyramidal feature hierarchy resulting in efficient object detection and recognition. SSD uses CNN on the image only once and creates a feature map. SSD uses the VGG-16 model pre-trained on MobileNet architecture for the extraction of the feature maps. The size of layers is not constant and this helps in detecting objects of various sizes.

Faster Region CNN FRCNN uses a CNN named Region Proposal Network to create the bounding boxes for the target object. FRCNN proves to faster than the Fast RCNN model by 10 times with similar accuracy. FRCNN proves to a model with high accuracy compared to other models. FRCNN constructs a single, unified model composed of RPN and fast RCNN with shared features of the convolutional layers.

Conclusion:

The FPGA-based implementation of object detection and recognition results in increased performance with the help of the PYNQ Z2 board and Movidius NCS. The YOLO model proves the fastest and easiest of the three models but has lesser accuracy and better FPS than FRCNN. The SSD has the highest FPS and also balances the computational time and accuracy.

FRCNN has the highest accuracy among all the models but with a larger computational time and the lowest FPS. The choice of object detection and recognition models depends on the application of the target.

The suitable object detection and recognition model can be selected based on the application required based on the results.

Related posts:

At the 1st of October the VipexCoin IEO will start, providing the opportunity to investors to participate in the creation of a great platform which will offer functions and services that only a…

The rest of us who try to enforce our boundaries without any kind of privilege or power to back us up are labelled as trouble makers who think to much of ourselves.

When we talk about the digital transformation of the workplace, we often focus on the business or technical aspects: supply chains, artificial intelligence, product services, and a whole host of…